Efficient Privacy-preserving Logistic Model With Malicious Security.

Area of Science:

• Cryptography and Data Security
• Machine Learning and Data Mining

Background:

• Secure computation is crucial for protecting data against malicious adversaries.
• Existing models often require multiple servers and an honest majority.
• Logistic regression is a widely used and effective classification model.

Purpose of the Study:

• To develop a novel, maliciously secure logistic regression model.
• To enable secure computation with a single, semi-honest server.
• To enhance privacy-preserving data mining techniques.

Main Methods:

• A new matrix encryption technique is proposed.
• The scheme utilizes a single semi-honest server.
• A lossy compression method minimizes communication costs.
• The $\mathcal{H}$ -transformation ensures indistinguishability against chosen-plaintext attacks.

Main Results:

• The proposed scheme is resilient to malicious data providers.
• Malicious activities are detectable during a verification stage.
• The method achieves accuracy comparable to non-private models.
• It demonstrates high efficiency for analyzing large-scale datasets.

Conclusions:

• The novel scheme provides efficient and accurate maliciously secure logistic regression.
• It outperforms existing frameworks in computational and communication costs.
• This work advances privacy-preserving data mining with a practical single-server solution.